{"id":"3d8e10b1-7804-4c65-bfb7-3e21ee9cbb51","arxiv_id":"2507.11561","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A multi-view VAE pretrained on newborn echocardiograms achieves higher balanced accuracy for PH severity grading than the supervised baseline, but falls behind it for binary PH detection.","lead":"This paper tests whether a multi-view variational autoencoder improves computer-based detection of pulmonary hypertension in newborn ultrasound videos. It finds the method mainly helps on severity grading, and does not beat a supervised baseline for simple yes/no detection.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 does not reproduce the headline severity advantage: MMVM-VAE ties Ind-VAE-FA on balanced accuracy, has lower AUROC than supervised, and no significance or patient-level clustering is reported; the central claim is therefore not established.","rationale":"The reader's conditional verdict is appropriate. I focused on the numerical support for the central claim rather than the surrogate-label issue, because a claim that fails internally cannot be rescued by better labels. Table 1 is cited as the main evidence for 'higher balanced accuracy and strong AUROC' in severity prediction, but the held-out test results in Table 3 show MMVM-VAE's balanced accuracy is identical to the independent-VAE feature-aggregation baseline (0.46) and its AUROC is lower than the supervised baseline (0.64 vs 0.72). The binary detection task, which the abstract also claims improves, is won by the supervised baseline. The generalization argument depends on comparing Table 2 (validation) with Table 3 (test), but both VAE variants exhibit the same drop in severity balanced accuracy (0.56→0.46), so the proposed multi-view mixture prior contributes no measurable robustness gain over simply concatenating independent VAE features. With three random seeds and no patient-level clustering or significance test, the Table 1 gap (0.56 vs 0.50) is within the reported standard deviations and may be seed noise. The label-surrogacy concern remains real but is secondary: PSAX-P defines the label and is also an input, so the task may be mimicking the cardiologist's visual reading; it should be addressed by independent label validation. The verdict should stay CONDITIONAL pending the permutation test and label validation.","tokens_in":11271,"tokens_out":8367,"duration_ms":86616,"concrete_test":"Run a patient-level clustered permutation test on the held-out test split (Table 3) comparing MMVM-VAE against Ind-VAE-FA and the supervised baseline for severity balanced accuracy and AUROC, with videos nested by newborn and the same random seeds. If the 95% CI for the MMVM−Ind-VAE-FA difference includes zero (or the permutation p-value is >0.05), the central claim reduces to 'competitive with independent-VAE pretraining' and Table 1's 0.56 vs 0.50 gap is not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that MMVM-VAE improves severity classification and generalization—rests on Table 1's balanced accuracy (0.56 vs 0.50 for Ind-VAE-FA and 0.42 for supervised), but the supplementary held-out test (Table 3) does not reproduce this advantage. On severity, MMVM-VAE and Ind-VAE-FA tie at 0.46±0.06 vs 0.46±0.04 balanced accuracy, and MMVM-VAE's AUROC (0.64±0.04) is below both Ind-VAE-FA (0.65±0.02) and the supervised baseline (0.72±0.05). For binary detection, the supervised baseline is strictly better (0.86/0.80/0.77 vs 0.78/0.75/0.75). The 'less decreasing in performance' claim also matches Ind-VAE-FA: both drop from 0.56 to 0.46 in severity balanced accuracy, whereas MMVM-VAE does not uniquely improve over independent VAE feature aggregation. No significance testing or patient-level clustering is reported, so the headline difference could reflect seed noise. Compounding this, the severity label is defined from the PSAX-P view, which is a model input, so even a significant gain would only establish mimicry of a single-view echocardiographic surrogate unless validated against RHC hemodynamics or an independent grading basis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage pipeline for predicting pulmonary hypertension (PH) in newborns from echocardiographic videos. In the first stage, a multi-view variational autoencoder (MMVM-VAE) with a mixture-of-experts prior is trained to reconstruct clips from five standard views; in the second stage, the learned latent representations are used to train a lightweight classifier for binary PH detection and three-class severity grading. The authors compare this approach against a supervised baseline (Ragnarsdottir et al., 2024) and against independently trained per-view VAEs with feature aggregation, reporting results on an in-house dataset that includes a held-out test cohort. The central claim is that the multi-view VAE representation improves generalization and severity classification accuracy relative to supervised learning.","tokens_in":11502,"tokens_out":6622,"duration_ms":67403,"significance":"The clinical problem is important and the dataset (936 videos from 192 newborns for training/validation, 375 videos from 78 newborns for held-out testing) is relatively large for neonatal echocardiography. The two-stage self-supervised approach is a plausible way to exploit unlabeled data, and the inclusion of a held-out test cohort is a genuine strength. The architecture is described clearly and the paper evaluates several baselines and multiple metrics. However, the evidence as reported does not support the headline claims: the held-out test table (Table 3) shows the proposed model tied with or worse than the baselines on key metrics, and the severity label is derived from an input view (PSAX-P), raising a circularity concern that is not addressed. These issues are load-bearing for the paper's central empirical claim.","major_comments":[{"comment":"The central claim that MMVM-VAE improves PH severity classification and generalization is not supported by the held-out test set. On severity prediction, MMVM-VAE achieves balanced accuracy 0.46±0.06, identical to Ind-VAE-FA (0.46±0.04) and only slightly above the supervised baseline (0.42±0.03), while its AUROC (0.64±0.04) is below both Ind-VAE-FA (0.65±0.02) and supervised (0.72±0.05). On binary detection, the supervised baseline is strictly better (AUROC 0.86 vs 0.78; balanced accuracy 0.77 vs 0.75). No significance tests are reported, so the small differences shown in Table 1 may reflect seed noise. The conclusion in §5 that the framework achieved 'more robust detection and grading' and 'less decreasing in performance' is therefore overstated.","section":"§4.5, Table 3"},{"comment":"The PH severity label is defined entirely from the PSAX-P view: 'PH grading was determined based on the PSAX-P view' with severity criteria based on septal flattening or bowing. Since PSAX-P is also one of the model inputs, a model can 'predict' the label by mimicking the cardiologist's qualitative reading of that same view. The paper does not discuss this circularity, and it does not validate the echocardiographic surrogate against right heart catheterization hemodynamics or an independent grading basis. As a result, the reported severity results cannot be interpreted as evidence that the model learns clinically meaningful PH severity beyond reproducing the PSAX-P-derived annotation.","section":"§4.1, §4.5"},{"comment":"The evaluation unit appears to be individual videos, but the data contain multiple videos per newborn (roughly 5 videos per patient in both cohorts). No patient-level clustering or aggregation is reported, and the 5-fold cross-validation described in Table 2 does not state whether splits are patient-disjoint. Metrics computed on videos from the same patient are not independent, so the reported standard deviations (across three seeds) underestimate variability and significance testing is missing. Please report patient-level metrics (e.g., per-patient majority voting) and patient-disjoint cross-validation details.","section":"§4.4, Tables 2–3"},{"comment":"The relationship between Tables 1, 2, and 3 is unclear and appears internally inconsistent. Table 1 reports MMVM-VAE multi-view severity balanced accuracy of 0.56±0.04 and AUROC 0.74±0.01, whereas Table 3 (held-out test) reports 0.46±0.06 and 0.64±0.04, and Table 2 (validation) reports 0.56±0.17 and 0.69±0.10. Similarly, the multi-view supervised numbers in Table 1 (0.86/0.80/0.77 for binary) match Table 3 but not Table 2. The text says the validation-to-test drop 'can't directly be observed in Table 1', yet Table 1 appears to duplicate Table 3 for some rows. Please specify which data split Table 1 reports and how it relates to Tables 2 and 3.","section":"§4.5, Tables 1–3"}],"minor_comments":[{"comment":"The caption contains a typo: 'Ind-VAE-FA and and MMVM-VAE' should read 'Ind-VAE-FA and MMVM-VAE'.","section":"Table 1 caption"},{"comment":"Several key hyperparameters are not reported, including latent dimension, optimizer, learning rate, batch size, number of training epochs, and the augmentation schedule used during VAE pretraining. These details are needed for reproducibility.","section":"§4.3"},{"comment":"Some table entries appear to be malformed or misprinted, e.g., '0.68 ±0.5' for PSAX-P MMVM-VAE balanced accuracy and several '±0.5' values that are inconsistent with the ranges elsewhere. Please proofread all tables carefully.","section":"Table 2"},{"comment":"The terms 'multi-view' and 'multi-modal' are used interchangeably (e.g., 'multi-view learning conceptually align with multi-modal learning'); the distinction and the intended mapping should be clarified.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a conference paper from the same research group that produced the baseline method (Ragnarsdottir et al., 2024) and the dataset. This is not by itself a flaw, but the editor may wish to verify that the novelty over that prior work is substantial. More importantly, the held-out test results in Table 3 do not support the main claims made in the abstract and conclusion; the authors should either revise the claims to match the evidence or provide additional analyses that substantiate the improvement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an incremental application of the authors' own MMVM-VAE to newborn echo for PH. The interesting claim—that multi-view VAE pretraining generalizes better to a held-out cohort—does not survive contact with their own Table 3.\n\nThe paper does some things right. It uses a real dataset with a held-out test set collected a year after training data, compares against a supervised baseline and independent VAEs, and puts per-view and multi-view results in the supplement. The writing is clear about the setup, and the limitations section is honest about single-center data and non-simultaneous views.\n\nThe central problem: the headline advantage appears only in Table 1's severity balanced accuracy (0.56 vs 0.50 for Ind-VAE-FA, 0.42 for supervised). On the held-out test in Table 3, MMVM-VAE ties Ind-VAE-FA at 0.46 balanced accuracy and is below supervised for AUROC (0.64 vs 0.72). For binary detection, supervised is strictly better in every metric. So the 'more robust detection and grading' claim is not supported. There are also no significance tests, videos from the same patient are treated as independent, and the severity label is defined visually from the PSAX-P view that is also a model input—so the model may just be mimicking the cardiologist's reading of that view.\n\nThese are fixable problems. The paper is not a fabrication or a mess; it just overstates what its own numbers show. The method is a reasonable extension of prior VAE work, and the application domain is clinically relevant.\n\nThis is useful reading for anyone working on self-supervised pretraining for medical video, and for clinicians who want to know the current state of automated PH screening. But I would not cite the generalization claim.\n\nWorth a serious referee. A good referee should ask for narrower claims, patient-level clustering, significance testing, and ideally an independent baseline or dataset. With those changes, it could be a solid contribution.","headline":"The generalization claim doesn't hold up in the paper's own held-out Table 3; still a legitimate application paper worth refereeing if claims are narrowed.","tokens_in":12142,"tokens_out":2235,"would_cite":false,"duration_ms":22439,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multi-view variational autoencoder with a data-dependent mixture-of-experts prior improves pulmonary hypertension severity classification in newborns and transfers to a held-out cohort more consistently than a supervised baseline.","keywords":["pulmonary hypertension","neonatal echocardiography","multi-view learning","variational autoencoder","representation learning","medical image classification","generalization"],"falsifier":"Remove the PSAX-P view from the multi-view input set and retrain the MMVM-VAE pipeline; if the severity balanced accuracy falls to the independent-VAE level on the held-out test set, the reported multi-view gain is explained by the model reading the label-defining view rather than by cross-view sharing.","tokens_in":11005,"feed_emoji":"🫀","tokens_out":6008,"duration_ms":59053,"temperature":0.7,"pith_summary":"The paper argues that a multi-view variational autoencoder can make automated pulmonary hypertension (PH) screening in newborns more reliable than the current supervised single- or multi-view approaches. It proposes a two-stage pipeline: first pretrain a multi-view VAE on echocardiography videos, with a data-dependent prior that lets each view's latent code borrow information from all other views, then fine-tune a lightweight classifier on the learned representations. On a held-out cohort of newborns from the same medical center, the method reaches a balanced accuracy of 0.56 and AUROC of 0.74 for three-class severity grading, compared with 0.42 and 0.72 for the supervised baseline; for binary detection it also shows smaller performance drops on test data. If true, this would make PH grading less dependent on a clinician's visual read and give an objective, scalable screening tool for newborns.","feed_headline":"Multi-view VAE beats supervised baseline on newborn PH grading","feed_subtitle":"Pretraining across five echo views keeps accuracy high on a held-out newborn cohort.","key_machinery":"The central object is the MMVM-VAE, a multi-view variational autoencoder with a variational mixture-of-experts prior. The prior is defined as $h(z \\mid X) = \\prod_{m=1}^{M} h(z_m \\mid X)$ with $h(z_m \\mid X) = \\frac{1}{M} \\sum_{\\tilde{m}=1}^{M} q_\\phi(z_m \\mid x_{\\tilde{m}})$, so each view's latent density is a mixture of all views' encodings. This soft-sharing aligns view-specific representations, captures both shared and view-specific features, and is what carries the generalization improvement on the held-out cohort.","core_discovery":"The central discovery is that forcing views to share a latent code through a variational mixture-of-experts prior improves the transferability of the learned features. In MMVM-VAE, the prior for each view is a mixture of the posterior encodings of all views, so the latent representation of any one view is influenced by the others. Pretraining this model on echocardiographic videos, including unlabeled clips, and then training a small classifier on the latents yields more robust detection and grading of PH on a held-out test cohort than training a supervised video classifier end-to-end. The paper reports stable results especially for the three-class severity task, where MMVM-VAE achieves higher balanced accuracy and a strong AUROC relative to alternatives (Table 1).","pith_inferences":["If the label-defining PSAX-P view is also a model input, a reader should suspect that at least part of the severity signal comes from that view; replacing or masking PSAX-P in an ablation would clarify whether the shared prior adds independent information.","The method assumes the standard views are acquired close enough in time to be meaningfully shared, but the videos are not synchronized, so cross-view sharing could underestimate the benefit of truly aligned multi-view data.","Should multi-center data become available, the same pretraining strategy can be tested on demographic shifts; a failure there would limit the generalization claim to within-center shifts.","The two-stage representation-learning pipeline may transfer to other neonatal echocardiographic tasks, such as predicting therapy response or adverse outcomes, which the paper names as future work."],"forward_implications":["PH severity can be graded automatically with higher balanced accuracy than an end-to-end supervised video classifier, so the screening step may require less of a cardiologist's time.","Because the VAE is pretrained on both labeled and unlabeled videos, routinely collected echocardiograms without labels can contribute to the model, easing the burden of building large annotated datasets.","The learned representation transfers more consistently to a held-out test cohort than the supervised baseline, suggesting the approach is better suited to deployment on data collected later or under slightly different conditions.","Multi-view inputs outperform single-view inputs across methods, confirming that the five standard echocardiographic views carry complementary information useful for PH assessment."],"supporting_citations":[{"why":"Supplies the dataset, the preprocessing pipeline, and the supervised multi-view baseline that the paper compares against.","marker":"Ragnarsdottir et al. (2024)"},{"why":"Defines the MMVM-VAE objective and the data-dependent variational mixture-of-experts prior that carries the method.","marker":"Sutter et al. (2024)"},{"why":"Provides the VAE formulation and evidence lower bound that underlies both the independent and multi-view models.","marker":"Kingma & Welling (2014)"},{"why":"Shows how aggregating independent posteriors into a joint posterior improves multi-modal representation learning, motivating the multi-view approach.","marker":"Wu & Goodman (2018)"},{"why":"Introduces the variational mixture-of-experts autoencoder that the MMVM-VAE prior builds on.","marker":"Shi et al. (2019)"},{"why":"Supplies the generalized multimodal ELBO that connects independent and joint posterior objectives.","marker":"Sutter et al. (2021)"},{"why":"Provides clinical evidence that multiple echocardiographic views improve PH assessment, motivating multi-view modeling.","marker":"Schneider et al. (2017)"},{"why":"Justifies the minority-class oversampling strategy used to handle class imbalance in training batches.","marker":"Buda et al. (2018)"}],"fun_headline_variants":["Multi-view VAE sharpens newborn PH severity grading","Shared latents improve PH prediction in newborns","Variational model outdoes supervised on newborn PH","Multi-view VAE edges out baselines for infant PH"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ground-truth PH severity labels come from one senior cardiologist's visual grading of septal flattening or bowing in the PSAX-P echo view, and the model also receives that view as an input, so the whole claim assumes that this visual surrogate is a valid severity measure and that using the label-defining view does not inflate apparent performance.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view VAE sharpens newborn PH severity grading","Shared latents improve PH prediction in newborns","Variational model outdoes supervised on newborn PH","Multi-view VAE edges out baselines for infant PH"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1428,"prompt_tokens":865,"completion_tokens":563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":502}},"tokens_in":481,"tokens_out":563,"duration_ms":6329,"temperature":1.0,"reasoning_tokens":502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:39:26.402292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove the PSAX-P view from the multi-view input set and retrain the MMVM-VAE pipeline; if the severity balanced accuracy falls to the independent-VAE level on the held-out test set, the reported multi-view gain is explained by the model reading the label-defining view rather than by cross-view sharing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dataset, the preprocessing pipeline, and the supervised multi-view baseline that the paper compares against."},{"cited_title":"Auto- Encoding Variational Bayes","cited_arxiv_id":null,"evidence_quote":"Provides the VAE formulation and evidence lower bound that underlies both the independent and multi-view models."},{"cited_title":"Siddharth, Brooks Paige, and Philip Torr","cited_arxiv_id":null,"evidence_quote":"Introduces the variational mixture-of-experts autoencoder that the MMVM-VAE prior builds on."},{"cited_title":"Lang, Gerald Maurer, Thomas Binder, and Georg Goliasch","cited_arxiv_id":null,"evidence_quote":"Provides clinical evidence that multiple echocardiographic views improve PH assessment, motivating multi-view modeling."},{"cited_title":"A systematic study of the class imbalance problem in convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Justifies the minority-class oversampling strategy used to handle class imbalance in training batches."}],"review_version":1}